---
source: 'https://howaiworks.ai/models/claude-fable'
section: models
title: Claude Fable 5.1
description: >-
  Anthropic's most capable widely released model: same $10/$50 pricing as Fable
  5, cache reads cut 75%, 1M context and always-on adaptive thinking.
tags:
  - Claude
  - Anthropic
  - Language Model
  - Large Language Model
  - AI Assistant
  - Frontier Model
  - Fable
  - Agentic AI
  - Latest
category: Language Models
developer: Anthropic
developerWebsite: 'https://www.anthropic.com/'
modelType: Language Model
releaseDate: '2026-09-01'
lastUpdated: '2026-09-06'
license: Proprietary
contextWindow: 1M tokens
inputPrice: $10
outputPrice: $50
knowledgeCutoff: June 2026
officialWebsite: 'https://www.anthropic.com/claude-fable-and-mythos-5-1'
docsPage: 'https://platform.claude.com/docs/en/models/fable-5-1/overview'
pricingPage: 'https://claude.com/pricing'
---

# Claude Fable 5.1

> Anthropic's most capable widely released model: same $10/$50 pricing as Fable 5, cache reads cut 75%, 1M context and always-on adaptive thinking.

## Overview

Claude Fable 5.1 is Anthropic's most capable widely released model, shipped on September 1, 2026 alongside the invitation-only Claude Mythos 5.1. Anthropic calls the pair "the world's most advanced models for coding and knowledge work."

It is a point release, and the interesting change is economic. Input and output prices are identical to [Claude Fable 5](https://howaiworks.ai/blog/claude-fable-5-announcement-2026) — $10 and $50 per million tokens — but a prompt cache read now costs $0.25 per million: 2.5% of the base input price, where every other Claude model charges 10%. For sessions that re-read a large cached prefix on every turn, cache reads are most of the bill, and Anthropic puts the saving at roughly 25% on a typical workload and up to 45% on highly agentic ones, measured over four weeks of its own August 2026 usage.

Fable 5.1 sits above the Opus tier rather than replacing it. Anthropic still recommends [Claude Opus 5](https://howaiworks.ai/models/claude-opus) as the starting point for most workloads, and positions Fable 5.1 for demanding reasoning and long-horizon agentic work "or when your evals on Claude Opus 5 at higher effort still fall short." Opus 5, [Sonnet 5](https://howaiworks.ai/models/claude-sonnet) and [Haiku 4.5](https://howaiworks.ai/models/claude-haiku) remain current in their tiers; there is no Opus 5.1 or Sonnet 5.1.

## Capabilities

Anthropic's own framing is that the gap over Fable 5 is widest at higher effort, and that at low or medium effort Fable 5.1 "achieves results similar to or better than Fable 5's at a much lower cost." The claimed gains concentrate in six areas:

- **Agentic coding over long sessions**: multi-file features, large refactors and migrations, debugging and code review across sessions that run for hours. Anthropic claims the model "avoids shortcuts that result in poorer-quality work" and fixes root causes rather than symptoms.
- **Knowledge work**: taking an analysis from a first question to a finished document, live-formula spreadsheet or slide deck.
- **Research and search**: higher accuracy on multistep web research that follows up on what it finds.
- **Vision**: dense charts, filings and tables nested in PDFs, including crop-and-zoom on charts.
- **Long-context work**: connecting details across the full 1M token [context window](https://howaiworks.ai/glossary/context-window).
- **Computer use**: operating a browser and desktop applications more reliably, and recovering from failed steps.

Multilingual performance is on par with Fable 5. The launch testimonials — Jane Street, Millennium — are vendor-selected and should be read as such.

## Technical Specifications

- **API model ID**: `claude-fable-5-1` (Bedrock: `anthropic.claude-fable-5-1`)
- **Context window**: 1M tokens, default and maximum (~555K words), at standard per-token pricing across the whole window. Same tokenizer as Fable 5 — roughly 30% more tokens for the same text than on models older than Opus 4.7.
- **Max output**: 128K tokens per synchronous request
- **Thinking**: [Adaptive thinking](https://howaiworks.ai/glossary/chain-of-thought) only, always on
- **Effort levels**: `low`, `medium`, `high`, `xhigh`, `max`. API default `high`; Claude Code defaults to high, Claude Cowork and Claude.ai to medium. Effort can be changed mid-conversation without invalidating the [prompt cache](https://howaiworks.ai/glossary/prompt-caching) (beta).
- **Training data cutoff**: June 2026 (reliable knowledge cutoff: June 2026) — five months later than Fable 5's January 2026
- **Latency**: Slower than Opus, Sonnet and Haiku, by Anthropic's own comparative rating
- **Data retention**: 30 days, required. A Covered Model; not available under zero data retention unless expressly authorized.
- **Retirement commitment**: not sooner than September 1, 2027

### Supported features

Effort, task budgets (beta), the memory tool, code execution, programmatic tool calling, context editing (beta), compaction, prompt caching and vision. New in 5.1: per-message effort, turn-scoped system messages that expire after one turn, and readable progress updates between tool calls via `thinking.display: "updates"` — all three in beta. Output carries Anthropic's statistical text watermark, with C2PA Content Credentials on generated media retrieved through the Files API.

## API Behavior

Fable 5.1 is not a drop-in swap for an Opus-tier model, and it is not quite a drop-in swap for Fable 5 either. Three changes are breaking:

- **Forced tool use is rejected.** `tool_choice` set to `{"type": "any"}` or `{"type": "tool", ...}` returns a 400. Thinking is always on, and a forced call would skip it. Use strict tool use or structured outputs instead.
- **Thinking blocks are bound to the model that produced them.** Fable 5.1 reads earlier models' blocks; no earlier model reads Fable 5.1's. A router that switches models mid-conversation silently drops them unless you opt into the reporting beta header.
- **Editing earlier turns invalidates thinking blocks.** Rebuilding the system prompt or tools array, injecting and then removing a per-turn reminder, or reordering messages produces a 400 on the next request for accounts created on or after August 31, 2026. Treat the conversation as append-only.

Carried over unchanged from Fable 5: adaptive thinking cannot be disabled; the raw chain of thought is never returned (`thinking.display` defaults to `"omitted"`, `"summarized"` is available); assistant prefill returns a 400; non-default `temperature`, `top_p` and `top_k` return a 400; and the minimum cacheable prompt is 512 tokens.

## Safety & Safeguards

Fable 5.1 keeps Fable 5's defining property: **it can decline a request**. The classifier categories are unchanged — `cyber`, `bio`, `frontier_llm` (the distillation classifier, which Anthropic warns benign machine learning work can trip) and `reasoning_extraction`, which fires when a request asks the model to reproduce its internal reasoning as text.

A declined request returns a **successful HTTP 200 response** with `stop_reason: "refusal"` and a `stop_details` object naming the policy area, not an error. Code that reads response content unconditionally will break.

Retries work three ways: server-side fallback via the `fallbacks` parameter (beta), SDK middleware, or your own retry logic. The permitted fallback targets are **Claude Opus 4.8 and Claude Opus 5**. You are not billed for a refusal that arrives before any output, and fallback credit refunds the prompt-cache cost of switching models.

## Claude Mythos 5.1

Claude Mythos 5.1 (`claude-mythos-5-1`) is the same underlying model as Fable 5.1 with some safeguards relaxed. It shares Fable 5.1's specifications, pricing and API behavior, and shipped the same day.

Access is invitation-only and, as of publication, restricted to vetted US organizations. The launch announcement describes it as "available to vetted cyberdefenders and life scientists" through two named channels: the **Cyber Verification Program**, which reduces cyber safeguards for defensive security work, and the **Life Sciences Verification Program**, an invite-only beta run in partnership with the US government that reduces biology safeguards for advanced life-sciences researchers. Anthropic states plainly that "currently, it is only available to a set of US organizations."

[Project Glasswing](https://anthropic.com/glasswing) has not been retired. Anthropic's model documentation still describes Mythos 5.1 as offered "by invitation only, as part of Project Glasswing," with access arranged through an Anthropic, AWS or Google Cloud account team, and Glasswing is itself being extended to roughly 150 more organizations in fifteen-plus countries. The verification programs are the route for cyber and life-sciences professionals; Glasswing remains the umbrella. Anthropic does not publish which safeguards remain active for each cohort, and we do not infer it.

## Prior Generation: Claude Fable 5 and the June 2026 Suspension

Claude Fable 5 (`claude-fable-5`) launched on June 9, 2026 as the first "Mythos-class model that we've made safe for general use," and remains available as a legacy model at the same $10/$50 pricing with the older $1 per million cache-read rate.

Its first month was not routine. On **June 12, 2026** the U.S. government imposed export controls on Fable 5 and Mythos 5, requiring access be restricted from foreign nationals inside or outside the United States. Anthropic could not verify nationality in real time and [suspended access to both models for all users](https://howaiworks.ai/blog/anthropic-suspends-fable-5-mythos-5-2026). The controls were lifted on **June 30, 2026** and access resumed on **July 1**. On redeployment Anthropic shipped an improved safety classifier targeting the bypass technique Amazon researchers had disclosed, reporting that it blocks the technique in over 99% of cases and redirects blocked requests to Claude Opus 4.8. It also restored Mythos 5 access for approved U.S. organizations and committed to pre-release testing and jailbreak information-sharing with the government.

Nothing in the 5.1 launch materials revisits that episode. It is worth remembering when planning capacity on a single frontier model: this one has been switched off before, by someone other than its vendor.

## Use Cases

- **Overnight autonomous engineering**: large-scale refactors and migrations that run unattended over hours, where the cache-read cut changes the arithmetic most.
- **Long-horizon research**: multi-hour investigations that gather context, build and self-verify before reporting.
- **Multi-agent orchestration**: coordinating long-running sub-agents, intervening only when one goes off track.
- **Enterprise deliverables**: financial analysis, live-formula spreadsheets, slide decks and documents produced end to end from a single brief.
- **Code review and debugging**: high-recall bug finding and correct identification of intermittent flakes — outside security-focused analysis, where the cyber classifiers apply.
- **Computer and browser use**: driving desktop applications over long task sequences, recovering from failed steps.

## Performance / Benchmarks

Anthropic published the 5.1 numbers **as a text table**, which is a change worth noting: the Fable 5 announcement in June carried its benchmark table only as an image, with no scores anywhere in the post text, and the figures had to be transcribed from the system card. That criticism no longer applies. The scores below are Anthropic's own, vendor-run, taken from the launch post.

| [Benchmark](https://howaiworks.ai/glossary/benchmark) | Claude Fable 5.1 | Claude Fable 5 | Claude Opus 5 |
|---|---|---|---|
| Terminal-Bench-Science 0.1 (agentic science) | 52.6% | 24.7% | 29.0% |
| Terminal-Bench 4.0 (agentic coding) | 55.8% | 42.0% | 52.3% |
| CursorBench 3.2.0 (agentic coding) | 73.4% | 70.5% | 70.0% |
| GDPval-AA v2 (knowledge work) | 1853 | 1723 | 1824 |
| OSWorld 2.0, partial credit (computer use) | 77.9% | 72.9% | 75.4% |
| OSWorld 2.0, strict (computer use) | 41.7% | 36.1% | 39.6% |
| Humanity's Last Exam, no tools | 60.9% | 57.8% | 56.6% |
| Humanity's Last Exam, with tools | 65.0% | 63.8% | 63.6% |
| AutomationBench (business workflows) | 31.4% | 17.1% | 26.9% |

Claude Mythos 5.1 scores **60.9%** on Terminal-Bench 4.0 against Fable 5.1's 55.8% — the clearest published measure of what the production safeguards cost.

Anthropic's footnote to the table is the part to read twice: "Fable 5.1 was evaluated with its production safeguards enabled. On tasks where these safeguards intervened, Fable 5.1 and Fable 5 scored zero on OSWorld 2.0, and Fable 5 scored zero on AutomationBench. In all other interventions from our safeguards, cybersecurity tasks were completed by Claude Opus 4.8, and biology tasks were completed by Claude Opus 5." Part of what these Fable rows measure is a different model finishing the job.

### Artificial Analysis

Fable 5.1's [Artificial Analysis page](https://artificialanalysis.ai/models/claude-fable-5-1) reports **57** on the Artificial Analysis Intelligence Index **v4.2**, for the configuration "Adaptive Reasoning, Max Effort, Default Fallback." Fable 5 scored 60 on **v4.1**. Those two numbers are not comparable: v4.2 changed the evaluation set, and scores never carry across index versions. The configuration string again names the fallback.

### How to read these numbers

These are **vendor-reported** figures except CursorBench, which Cursor measures. Anthropic chose the benchmarks, the harnesses and the effort levels, ran the evaluations itself, and evaluated Fable with the safeguards that can hand a trajectory to another Claude model mid-run. Its published comparison against OpenAI's [GPT-5.6 Sol](https://howaiworks.ai/models/gpt) uses the same harness. Treat the table as an upper bound and validate on your own set — particularly at the low and medium effort levels, where Anthropic's cost claims live and its published numbers do not.

## Limitations

- **Cost**: $10/$50 per million tokens still makes this Anthropic's most expensive generally available model, double the Opus tier. The cache-read cut helps only workloads that actually re-read a cached prefix; a single-shot request pays exactly what Fable 5 paid.
- **Latency**: Rated "slower" by Anthropic's own comparison, and a hard task at high effort can occupy one request for many minutes.
- **Refusals**: Four classifier categories can decline a request, including false positives on legitimate adjacent work. Integrations must handle the `refusal` stop reason and should configure a fallback.
- **No zero data retention**: the 30-day requirement rules it out for organizations with ZDR configured.
- **Reduced API surface**: no thinking-disabled mode, no sampling parameters, no assistant prefill, no forced tool use, and no raw chain of thought.
- **Migration is not free**: thinking-block binding means an integration that edits conversation history starts failing on requests that worked against Fable 5.
- **Behavioral drift from Fable 5**: fewer batched tool calls, fewer progress updates, less tool use at `low` effort, denser prose, and whole-file rewrites where a targeted edit would do.
- **No internet access**: cannot browse unless integrated through tool use or web search.

## Pricing & Access

### API Pricing

Available via the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud and Microsoft Foundry:

| Item | Claude Fable 5.1 | Claude Fable 5 |
|---|---|---|
| Input | $10 / MTok | $10 / MTok |
| Output | $50 / MTok | $50 / MTok |
| 5-minute cache write | $12.50 / MTok | $12.50 / MTok |
| 1-hour cache write | $20 / MTok | $20 / MTok |
| Cache read | **$0.25 / MTok** | $1 / MTok |

The cache read line is the whole release. Fable 5.1 and Mythos 5.1 price cache hits at 0.025x base input; every other Claude model charges 0.1x. A long session that re-reads a warm prefix pays a quarter of the Fable 5 rate, which is where the "roughly 25% cheaper, up to 45% for highly agentic work" claim comes from — derived from Anthropic's own August 2026 traffic, not an independent study. Cache writes and the 512-token minimum cacheable prompt are unchanged, so the break-even point moves in favor of caching almost everything.

Batch processing is $5 / $25 per million tokens. US-only inference via `inference_geo` adds a 1.1x multiplier on every token category.

### Consumer Plans

Fable 5.1 is available on Claude.ai, Claude Code and Claude Cowork. Effort defaults differ by surface: high in Claude Code, medium in Claude Cowork and on Claude.ai. Usage credits apply beyond the included allowance.

## Ecosystem & Tools

### Official SDKs
- **[Python SDK](https://github.com/anthropics/anthropic-sdk-python)**: async support
- **[TypeScript SDK](https://github.com/anthropics/anthropic-sdk-typescript)**: Node.js and browsers
- **REST API**: direct HTTP access from any language

Refusal-fallback middleware ships in the TypeScript, Python, Go, Java and C# SDKs.

### Developer Tools
- **[Anthropic Console](https://console.anthropic.com/)**: playground for testing and prompt development
- **[Claude Code](https://claude.com/product/claude-code)**: agentic coding in the terminal, IDE and web
- **[Claude.ai](https://claude.ai)**: browser-based chat

## Community & Resources

- [Official Announcement: Claude Fable 5.1 and Claude Mythos 5.1](https://www.anthropic.com/claude-fable-and-mythos-5-1)
- [Claude Fable 5.1 model documentation](https://platform.claude.com/docs/en/models/fable-5-1/overview)
- [What's new in Claude Fable 5.1](https://platform.claude.com/docs/en/models/fable-5-1/whats-new-fable-5-1) - Breaking changes, new features and migration steps
- [Claude Fable 5.1 and Claude Mythos 5.1 System Card](https://www.anthropic.com/claude-fable-5-1-mythos-5-1-system-card)
- [Claude Mythos](https://www.anthropic.com/claude/mythos) - Verification programs and access
- [Project Glasswing](https://anthropic.com/glasswing)
- [Redeploying Claude Fable 5](https://www.anthropic.com/news/redeploying-fable-5) - Statement on the June 2026 export-control suspension and restoration
- [Pricing Page](https://claude.com/pricing)

## Frequently Asked Questions

### When was Claude Fable 5.1 released?

Claude Fable 5.1 was released on September 1, 2026, alongside the invitation-only Claude Mythos 5.1. It succeeds Claude Fable 5, which shipped on June 9, 2026 and remains available as a legacy model.

### What is the model ID for Claude Fable 5.1?

The Claude API model ID is `claude-fable-5-1`, and the same string is used on Google Cloud, Microsoft Foundry and Claude Platform on AWS. On Amazon Bedrock it carries the `anthropic.` prefix: `anthropic.claude-fable-5-1`.

### How much does Claude Fable 5.1 cost?

Base pricing is unchanged from Fable 5 at $10 per million input tokens and $50 per million output tokens. What changed is prompt caching: a cache read costs $0.25 per million tokens, 2.5% of the base input price instead of the usual 10% — a 75% cut. Anthropic says that lowers a typical bill by roughly 25%, and up to 45% for highly agentic workloads, based on its own August 2026 usage.

### What is the context window for Claude Fable 5.1?

1M tokens, the default and the maximum, at standard per-token pricing across the whole window, with up to 128K output tokens per synchronous request. It uses the tokenizer introduced with Claude Opus 4.7, so the same text produces roughly 30% more tokens than on models older than Opus 4.7.

### What effort levels does Claude Fable 5.1 support?

Five: `low`, `medium`, `high`, `xhigh` and `max`. The Claude API default is `high`. In Anthropic's own products the defaults differ — high effort in Claude Code, medium in Claude Cowork and on Claude.ai. Fable 5.1 can also change effort mid-conversation without invalidating the prompt cache, in beta.

### How is Claude Fable 5.1 different from Claude Opus 5?

Fable 5.1 costs twice as much per token and is slower, in exchange for depth on the hardest reasoning and long-horizon agentic work. It also behaves differently: adaptive thinking is always on and cannot be disabled, the raw chain of thought is never returned, sampling parameters and assistant prefill are rejected, forced tool use returns an error, and safety classifiers can decline a request. Anthropic still recommends Opus 5 as the starting point for most workloads.

### What is Claude Mythos 5.1 and how do I get access?

Claude Mythos 5.1 (`claude-mythos-5-1`) is the same underlying model as Fable 5.1 with more permissive safeguards for professional cybersecurity and life-sciences work. It is invitation-only. Anthropic's announcement routes access through the Cyber Verification Program and the Life Sciences Verification Program, currently limited to a set of vetted US organizations; the model docs describe it as offered to Project Glasswing participants, with access arranged through an Anthropic, AWS or Google Cloud account team. Customers without access should use Fable 5.1, which has the same capabilities.

### Why does Claude Fable 5.1 sometimes refuse a request?

It ships with safety classifiers covering the same categories as Fable 5: `cyber`, `bio`, `frontier_llm` (the distillation classifier, which Anthropic warns benign machine learning work can trip) and `reasoning_extraction`. A declined request returns a successful HTTP 200 response with `stop_reason: "refusal"` rather than an error, so check the stop reason before reading the content. Permitted fallback targets are Claude Opus 4.8 and Claude Opus 5.

### Can I turn off thinking on Claude Fable 5.1?

No. Adaptive thinking is the only thinking mode and it is always on. Both `thinking: {"type": "disabled"}` and the older `"enabled"` plus `budget_tokens` form return a 400 error. Use the effort parameter to control reasoning depth and token spend instead.

### What breaks when migrating from Claude Fable 5 to Claude Fable 5.1?

Three changes are breaking. Forced tool use (`tool_choice` of `any` or `tool`) returns a 400 error. Thinking blocks are bound to the model that produced them, so earlier models cannot read Fable 5.1's. And editing anything before a thinking block — a rebuilt system prompt, an injected per-turn reminder, a reordered message — invalidates every later block. Keep the conversation append-only and re-tune effort from the new default.

### Does Claude Fable 5.1 work under zero data retention?

No. Fable 5.1 and Mythos 5.1 are designated Covered Models and carry a 30-day data retention requirement for safety monitoring. They are not available under zero data retention unless Anthropic expressly authorizes it.

## Related

### Related models

- [Claude Opus 5](https://howaiworks.ai/models/claude-opus)
- [Claude Sonnet 5](https://howaiworks.ai/models/claude-sonnet)
- [Claude Haiku 4.5](https://howaiworks.ai/models/claude-haiku)
- [GPT-6 Astra](https://howaiworks.ai/models/gpt)
- [Gemini 3.8 Flash](https://howaiworks.ai/models/gemini)
- [Grok 4.6](https://howaiworks.ai/models/grok)

---

Source: https://howaiworks.ai/models/claude-fable — HowAIWorks.ai